Thesis: Ph. D., Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, 2014. / Cataloged from PDF version of thesis. / Includes bibliographical references (pages 122-138). / The invariance hypothesis is the claim that properties of the ventral stream can be deduced from 1. a consideration of the computational problem with which it is engaged: invariant object recognition, 2. constraints of the neural "hardare", and 3. the visual environment. We are interested here in a particular instantiation of this idea. A recent general theory of hierarchical networks for invariant recognition [1] describes many modern convolutional networks as special cases, and also implies the existence of a wider class of algorithms, which we are only now beginning to explore. Our version of the invariance hypothesis is the claim that the algorithm implemented by the ventral stream is also in this class. As applied to the brain, the theory follows from a few simple and commonly accepted premises. This thesis contributes several models/studies in which properties of the ventral stream are deduced and explained in the context of the theory. The main contribution here is providing a general framework through which disparate results concerning many parts of the ventral stream, and even different levels of analysis [2], can be bridged and understood. In that sense, it is primarily a Neuroscience contribution. However, the ideas and algorithms it suggests may also have implications for the broader question of how to learn representations capable of supporting intelligence. / by Joel Zaidspiner Leibo. / Ph. D.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 2012. / Cataloged from PDF version of thesis. / Includes bibliographical references (p. 127-139). / This thesis proposes a computational framework for understanding human Theory of Mind (ToM): our conception of others' mental states, how they relate to the world, and how they cause behavior. Humans use ToM to predict others' actions, given their mental states, but also to do the reverse: attribute mental states - beliefs, desires, intentions, knowledge, goals, preferences, emotions, and other thoughts - to explain others' behavior. The goal of this thesis is to provide a formal account of the knowledge and mechanisms that support these judgments. The thesis will argue for three central claims about human ToM. First, ToM is constructed around probabilistic, causal models of how agents' beliefs, desires and goals interact with their situation and perspective (which can differ from our own) to produce behavior. Second, the core content of ToM can be formalized using context-specific models of approximately rational planning, such as Markov decision processes (MDPs), partially observable MDPs (POMDPs), and Markov games. ToM reasoning will be formalized as rational probabilistic inference over these models of intentional (inter)action, termed Bayesian Theory of Mind (BToM). Third, hypotheses about the structure and content of ToM can be tested through a combination of computational modeling and behavioral experiments. An experimental paradigm for eliciting fine-grained ToM judgments will be proposed, based on comparing human inferences about the mental states and behavior of agents moving within simple two-dimensional scenarios with the inferences predicted by computational models. Three sets of experiments will be presented, investigating models of human goal inference (Chapter 2), joint belief-desire inference (Chapter 3), and inference of interactively-defined goals, such as chasing and fleeing (Chapter 4). BToM, as well as a selection of prominent alternative proposals from the social perception literature will be evaluated by their quantitative fit to behavioral data. Across the present experiments, the high accuracy of BToM, and its performance relative to alternative models, will demonstrate the difficulty of capturing human social judgments, and the success of BToM in meeting this challenge. / by Chris L. Baker. / Ph.D.

Thesis: Ph. D., Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, 2016. / Cataloged from PDF version of thesis. / Includes bibliographical references (pages 205-211). / A number of recent proposals have used techniques from game theory and Bayesian cognitive science to formalize Gricean pragmatic reasoning [29, 28, 36, 51]. In the first part of this work, we discuss several phenomena which pose a challenge to these accounts of pragmatics: M-implicatures [45] and embedded implicatures which violate Hurford's constraint [49, 16]. While techniques have been developed for deriving M-implicatures, Hurford-violating embedded implicatures pose a more fundamental challenge to the models' architecture. In order to explain these phenomena, we propose that the semantic content of an utterance is not fixed independent of pragmatic inference; rather, pragmatic inference partially determines an utterance's semantic content. We show how semantic inference can be realized as an extension to the Rational Speech Acts framework [36]. The addition of lexical uncertainty derives both M-implicatures and the relevant embedded implicatures. This principle explains a novel class of implicature, non-convex disjunctive implicatures. These implicatures can be preserved in downward-entailing contexts in the absence of accenting, a property which is predicted by lexical uncertainty, but which violates prior generalizations in the literature [46, 27] In the second part of the thesis, we combine these pragmatic models with another recent probabilistic approach to natural language understanding, exploring the formal pragmatics of communication on a noisy channel. We extend a model of rational communication between a speaker and listener, to allow for the possibility that messages are corrupted by noise. Prosodic stress is modeled as the choice to intentionally reduce the noise rate on a word. We show that the model derives several well-known changes in meaning associated with stress, including exhaustive interpretations, scalar implicature strengthening, the association between stress and disagreement, and the interpretation of the focus-sensitive adverbs. We then show that it can account for several phenomena which are outside of the scope of previous accounts of stress interpretation: the effects of stress on quantifier domain inferences, the intensification of gradable adjective interpretation, and the strengthening of hyperbolic utterances. The account avoids the use of syntactic or semantic representations of stress; the interpretive effects of stress are derived from general-purpose pragmatic reasoning. / by Leon Bergen. / Ph. D.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 2004. / Includes bibliographical references (p. 120-125). / Over the last two decades numerous examples have demonstrated the remarkable plasticity of the developing brain. This plasticity occurs from the level of a single synapse to the repatterning of sensory input. One paradigm that demonstrates this plasticity is the re-routing of sensory input to inappropriate targets. This cross-modal plasticity in an animal model is reminiscent of similar rearrangements in deaf and blind human patients. In these animal models, visual input is induced to innervate the auditory or somatosensory thalamus, MGN and VB respectively, as a result of deafferentation of these nuclei. Such experiments have demonstrated that structures are influenced by their input, and therefore sensory input is able to use alternative pathways for function. This thesis examines the extent to which cues intrinsic to the target provide information to these novel retino-MGN projections. It will consider two examples in which the target structure imposes order onto the incoming sensory input; via intra-nuclei patterning and via a behaviorally relevant efferent pathway. We demonstrate that retinal axons use an ephrin gradient present in the MGN to acquire orderly connections, akin to retinal patterning in visual targets. Using fear conditioning, we show that learning of a visual cue changes when visual input is routed through the auditory pathway. To better understand the intrinsic cues present in a target, we identify a set of genes differentially expressed in the LGN and MGN, which includes a list of transcription factors and putative downstream targets. / (cont.) Furthermore, we demonstrate that deafferentation of the MGN does not influence these sensory-specific molecular profiles but does create a permissive environment which induces innervation by local axons. / by Charlene Ellsworth. / Ph.D.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 1993. / Includes bibliographical references (153-164). / by Amnon Shashua. / Ph.D.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, February 2010. / Cataloged from PDF version of thesis. / Includes bibliographical references (p. 114-123). / A central feature of visual pathway development is its organization into retinotopic maps. The developmental process by which these maps form involves a transition from early patterning cues to arrays of axonal guidance factors allowing the relative order of retinotopic axons to be preserved. Mechanisms linking patterning molecules of early development to topographic wiring and subsequent functional responses are not well understood. In this thesis, I performed a microarray screen comparing gene expression in early visual and auditory regions of the thalamus in order to identify early patterning candidates with a potential role in visual pathway differentiation. Among the candidates enriched in the visual thalamus, the transcription factor, Zic4, was found to be expressed in gradients of the developing retina, lateral geniculate nucleus (LGN) and primary visual cortex (V 1). Mice lacking Zic4 exhibited a deficit in eye-specific patterning to the thalamus that was complementary to the phenotype seen in mice lacking Tenm3, a type II homophilic transmembrane receptor and transcriptional regulator. Using intrinsic signal optical imaging techniques, I characterized the functional properties of primary visual cortical retinotopic maps in Zic4 and Ten_m3 null mice and identified complementary changes in the ipsilateral representation of V1, as well as evidence for eye-specific mismatch in the cortical binocular zone. Additionally, complementary positional shifts in VI were found in these mutants identifying a bidirectional modulation of mapping mechanisms in the visual pathway. / (cont.) In order to test whether Zic4 and Ten_m3 interact in serial or parallel pathways, I analyzed the retinogeniculate and cortical maps in the combination mutant. The Ten_m3/Zic4 double null mouse exhibited a partial rescue of retinogeniculate mapping and a complete reversal of the cortical changes found in either mutant alone, suggesting that the two genes interact to modulate common downstream effectors in opposite directions. In sum, this thesis presents a gene microarray screen used to identify Zic4 as a novel visual patterning gene, characterizes its loss-of-function phenotype on retinotopic mapping in the thalamus and cortex, and studies its antagonistic interaction with Ten_m3, another visual pathway patterning gene with a complementary loss-of-function phenotype. / by Sam H. Horng. / Ph.D.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 1995. / Includes bibliographical references. / by Jonathan S. Wallach. / Ph.D.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Brain and Cognitive Sciences, 2009. / Includes bibliographical references (p. 137-144). / In this thesis, we present three image estimation and enhancement algorithms inspired by human vision. In the first part of the thesis, we propose an algorithm for mapping one image to another based on the statistics of a training set. Many vision problems can be cast as image mapping problems, such as, estimating reflectance from luminance, estimating shape from shading, separating signal and noise, etc. Such problems are typically under-constrained, and yet humans are remarkably good at solving them. Classic computational theories about the ability of the human visual system to solve such under-constrained problems attribute this feat to the use of some intuitive regularities of the world, e.g., surfaces tend to be piecewise constant. In recent years, there has been considerable interest in deriving more sophisticated statistical constraints from natural images, but because of the high-dimensional nature of images, representing and utilizing the learned models remains a challenge. Our techniques produce models that are very easy to store and to query. We show these techniques to be effective for a number of applications: removing noise from images, estimating a sharp image from a blurry one, decomposing an image into reflectance and illumination, and interpreting lightness illusions. In the second part of the thesis, we present an algorithm for compressing the dynamic range of an image while retaining important visual detail. The human visual system confronts a serious challenge with dynamic range, in that the physical world has an extremely high dynamic range, while neurons have low dynamic ranges. / (cont.) The human visual system performs dynamic range compression by applying automatic gain control, in both the retina and the visual cortex. Taking inspiration from that, we designed techniques that involve multi-scale subband transforms and smooth gain control on subband coefficients, and resemble the contrast gain control mechanism in the visual cortex. We show our techniques to be successful in producing dynamic-range-compressed images without compromising the visibility of detail or introducing artifacts. We also show that the techniques can be adapted for the related problem of "companding", in which a high dynamic range image is converted to a low dynamic range image and saved using fewer bits, and later expanded back to high dynamic range with minimal loss of visual quality. In the third part of the thesis, we propose a technique that enables a user to easily localize image and video editing by drawing a small number of rough scribbles. Image segmentation, usually treated as an unsupervised clustering problem, is extremely difficult to solve. With a minimal degree of user supervision, however, we are able to generate selection masks with good quality. Our technique learns a classifier using the user-scribbled pixels as training examples, and uses the classifier to classify the rest of the pixels into distinct classes. It then uses the classification results as per-pixel data terms, combines them with a smoothness term that respects color discontinuities, and generates better results than state-of-art algorithms for interactive segmentation. / by Yuanzhen Li. / Ph.D.

Thesis: Ph. D. in Cognitive Science, Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, 2017. / Cataloged from PDF version of thesis. / Includes bibliographical references (pages 189-211). / I explore the hypothesis that the universal properties of human languages can be explained in terms of efficient communication given fixed human information processing constraints. I argue that under short-term memory constraints, optimal languages should exhibit information locality: words that depend on each other, both in their interpretation and in their statistical distribution, should be close to each other in linear order. The informationtheoretic approach to natural language motivates a study of quantitative syntax in Chapter 2, focusing on word order flexibility. In Chapter 3, I show comprehensive corpus evidence from over 40 languages that word order in grammar and usage is shaped by working memory constraints in the form of dependency locality: a pressure for syntactically linked words to be close. In Chapter 4, I develop a new formal model of language processing cost, called noisy-context surprisal, based on rational inference over noisy memory representations. This model unifies surprisal and memory effects and derives dependency locality effects as a subset of information locality effects. I show that the new processing model also resolves a long-standing paradox in the psycholinguistic literature, structural forgetting, where the effects of memory appear to be language-dependent. In the conclusion I discuss connections to probabilistic grammars, endocentricity, duality of patterning, incremental planning, and deep reinforcement learning. / by Richard Landy Jones Futrell. / Ph. D. in Cognitive Science

Thesis: Ph. D., Massachusetts Institute of Technology, Department of Brain and Cognitive Sciences, 2018. / Cataloged from PDF version of thesis. / Includes bibliographical references (pages 188-214). / Research in the domain of cognitive science has tended to neglect emotions. In my thesis, I take several steps to fill this gap by looking at people's representation of emotions, and its connection to other representations typically studied in cognitive science. I argue that people have an intuitive theory of emotion that is causally intertwined with their understanding of the physical and social world broadly. This intuitive theory allows us to use observed emotional cues as a window, to recover unobserved information about the world. I study these abilities in both adults and children, to gain insight into the most fundamental representations supporting such abilities. I also use computational models to capture the hierarchical, causal structure of this intuitive theory of emotion. In Study 1, I show that infants as young as 12-17 months can discriminate diverse within-valence emotional expressions elicited by funny, exciting, adorable, delicious, and sympathetic events, and map them onto their probable causes. In Study 2.1, I present that preschoolers can recover rich mental state information from observed emotional expressions. When the valence of someone's face changes between anticipated and actual outcomes, children by five gain insight into what she wants and believes about the world. Study 2.2 bridges theory of mind research, accounts of emotion attribution, and formal modeling, to provide a formal account of how people jointly infer beliefs and desires from emotional expressions. Study 3 tests children's understanding of social display rules. By middle childhood, children can use one person's emotional expressions regulated by a social context to infer the mental states of another. Altogether, these findings suggest that emotional cues provide a valuable entrée into the unseen world. Not only adults, but also children, can use observed emotional expressions to infer their external causes and the internal mental states of other people. Although this intuitive theory of emotion may not necessarily mirror the actual processes of how emotions are generated, it supports rational inferences much of time, and it may be formed early in development. I see this work as bridging gaps across disciplines and helping advance the cognitive science of emotion understanding. / "Funding support, including Leventhal, Stark, and Henry E. Singleton Fellowships, and the Center for Brains, Minds and Machines (funded by NSF STC Award CCF-1231216)"--Page 5 / by Yang Wu. / Ph. D.